0:00

We just launched this experiment and then

0:02

we are very surprised to see that

0:04

the hugely overparameterized model not only train

0:06

out of the box, like you have

0:09

very nice training curves, but also they

0:11

don't overfit aggressively at all. And what

0:13

we found empirically is that we just

0:15

out of the box, use typical like

0:18

supervised training. We don't have to play

0:20

with the hyperparameter optimizer, and you have

0:22

very very stable training. So

0:26

this brings us to

0:28

the question, is it

0:30

worth it to spend

0:32

so much money to

0:34

gather gigantic pre-training data

0:36

set, spend like months

0:38

on many GPUs to

0:40

produce those models, but

0:42

at least for some

0:44

application it seems to not be

0:46

much better than from there. based in

0:49

Switzerland. They have an amazing team. You've

0:51

seen many of the folks on the

0:53

team. They acquired Mine's AI, of course.

0:55

They did a lot of great work

0:57

on arc. They're now working on O1-style

0:59

models and reasoning and thinking and test

1:01

time computation. The reason you want to

1:03

work for them is you get loads

1:05

of autonomy, you get visibility, you can

1:07

publish your research, and also they are

1:10

hiring as well as ML engineers, the

1:12

hiring a chief scientist. They really, really

1:14

want to find the best possible person

1:16

for this role for this role. as

1:18

a joining bonus. So if you're interested

1:20

in working for them as an MO

1:22

engineer or their chief scientist, get in

1:24

touch with Benjamin Cruzier, go to two

1:26

for labs.a.i. and see what happens. Originally,

1:29

the main motivation was to see,

1:31

okay, how much information you gain

1:34

by doing play training, right? And

1:36

is this next second prediction really

1:38

making your network learn something about

1:40

language and reasoning? And so then

1:42

we are saying, one way to

1:45

compare this at least empirically is

1:47

to just take a randomly initialized

1:49

model. train it from scratch on

1:51

a supervised task like sentiment prediction,

1:53

sentiment analysis. And then in theory,

1:56

because we have a very very

1:58

small training data, let's say... samples

2:00

and because of Smodel, I've like... 7

2:02

billion parameters. The pre-train one will perform

2:05

very nicely with a little bit of

2:07

lower-off and tuning because it already knows

2:09

how to reason about the world, right?

2:11

So maybe you just adjust a little

2:13

bit to the specific task that you

2:16

want, but since you have so many

2:18

prior knowledge, you will serve the task

2:20

very easily. But the random one, either

2:22

will over-feed completely. Because you have like

2:24

7 billion parameters and only 20,000 training

2:27

samples. Or maybe it will not learn

2:29

at all because training dynamics will be

2:31

completely. So we just launched this experiment

2:33

and then we were very surprised to

2:35

see that the 7 billion, or like

2:37

the hugely overparameterized model, not only train

2:40

out of the box, like you have

2:42

very nice training curves, almost like you

2:44

train the EMIST, but also they don't

2:46

overfit aggressively at all. Like they overfit

2:48

less than if you just train a

2:51

MLP on EMIS, basically. And this is

2:53

very surprising. And so, basically, from this,

2:55

we said, okay, actually, maybe there is

2:57

a deeper question. which could be how

2:59

much implicit bias you have in those

3:02

language model because already we knew from

3:04

computer visions that for example Imagineet you

3:06

can have a 50 million model on

3:08

a 1 million data set so you

3:10

have this 50 to 1 ratio and

3:13

you have the implicit bias that prevents

3:15

you from overfitting and just serving the

3:17

task right but still it's 50 to

3:19

1 so this may sound a lot

3:21

for you know like. statistician, but now

3:23

it's like 7 billion to 20,000. So

3:26

like the ratio is like gigantic, right?

3:28

And yeah, to me it was very

3:30

surprising that the scale, like the size

3:32

of this ratio, still... allows you to

3:34

learn some things that does not overfit.

3:37

This is very surprising because in vision,

3:39

for example, transformers are known to overfit

3:41

more resilience and resonate. So they seem

3:43

at least in vision to have actually

3:45

less implicit bias or implicit regularization, but

3:48

at least with this. type of next

3:50

token, causal architecture or LLLM. Yeah, you

3:52

don't seem to work easily to your

3:54

data. So this was quite surprising. Yeah,

3:56

and we should bring in the name.

3:58

So this was your workshop paper at

4:01

the self-supervised learning workshop here at New

4:03

York State School. Correct. For perception tasks,

4:05

is LLLM pre-training by next token prediction

4:07

worth the cost? So this is absolutely

4:09

fascinating, right? So we've been given this,

4:12

this belief that we need to have

4:14

these huge... pre-trained models, they're trained on

4:16

all the data on the internet, and

4:18

it turns out that certainly for discrimination

4:20

tasks, so things like classification, urban generation,

4:23

actually you can just start from scratch

4:25

with a fairly small model and you

4:27

get sometimes even better results. Yeah, yeah,

4:29

and even small or even large model,

4:31

like you just start from scratch, you

4:34

do this very simple supervised classification task,

4:36

right? Okay, given this prompt, is it

4:38

good or bad? sentiment or what type

4:40

of job is the problem describing. You

4:42

know, this type of, I will not

4:44

call it reasoning, but you have more

4:47

semantic classification and turns out that you

4:49

start from random. Even if you have

4:51

a small training data set, we'll have

4:53

performances that are sometimes as good than

4:55

a pre-train model. So this brings us

4:58

to the question, is it worth it

5:00

to spend so much money to gather

5:02

gigantic pre-training data set, spend like months

5:04

on many GPUs to produce those models?

5:06

And for some cases, so for generation...

5:09

Alright, there is no question this is

5:11

what you need to do. You have

5:13

your next second prediction, you learn how

5:15

to generate samples, but at least for

5:17

some application it seems to not be

5:19

much better than random. So it's quite

5:22

interesting. So what are the differences in

5:24

the land representations? So that's something we

5:26

did not really look at like low

5:28

dimensional representation for what you learn. It's

5:30

possible, so some work, try to look

5:33

at the attention, entropy and the like

5:35

you... those mechanistic interpretability viewpoint of LLLM's.

5:37

So it would be interesting to see

5:39

if you have this sort of non-normal

5:41

collapse thing that happens. So even if

5:44

you're like a 7 billion parameter, maybe

5:46

you end up learning a very, very,

5:48

very simple sub-network that does the task.

5:50

It will be like lotoreticate hypothesis as

5:52

well. And that naturally emerged from the

5:55

training dynamics. Or is it really exploiting

5:57

all the parameters? I think that's one

5:59

thing. So to extend the workshop paper

6:01

to conference we want to probe into.

6:03

more product. What are the useful parameters?

6:05

What did they learn? Are each layer

6:08

actually learning something or maybe the first

6:10

layers don't really learn anything just the

6:12

last few ones learning something? So yes,

6:14

I was like... lots of open questions

6:16

for this. What does it tell us

6:19

about the nature of understanding and maybe

6:21

even intelligence because we think that the

6:23

reason why these things understand is because

6:25

they just have all of these representations

6:27

to all of these different you know

6:30

things in their experience. And and now

6:32

we can shortcut to you know to

6:34

want the better word. What does that

6:36

tell us? Yeah I think that's a

6:38

good question. So in this case we

6:40

must look at very specific classification tasks.

6:43

So for example of a job or

6:45

job. is it like a good or

6:47

bad sentiment? And this you are able

6:49

to solve it good, but you are

6:51

not able to go out of distribution

6:54

to solve a new type of question.

6:56

For example, for this job description, then

6:58

you cannot answer, okay, is this job

7:00

paying you more than this job? Because

7:02

this was not present in the training

7:05

data, right? So I think you get

7:07

very good models, cheaply. quickly from a

7:09

domination, but they will be very specialized.

7:11

And I think the benefit of having

7:13

maybe the pre-training may come if you

7:16

want to do more of like open-ended

7:18

classification or reasoning. So I think it

7:20

really depends on the type of application

7:22

you want to solve, what's your downstream

7:24

task, and how much you want to

7:26

generalize to new scenarios. But at least

7:29

now it shows that it's not just

7:31

pre-training with next second prediction is better

7:33

for everything. five years data scientists used

7:35

to build. specific models for doing everything.

7:37

And now we're in this regime of,

7:40

we need these really big models and

7:42

we do in context learning and maybe

7:44

even some fine tuning and we get

7:46

them to do fairly specific discriminative tasks.

7:48

But now you're saying we should almost

7:51

go back to where we were five

7:53

years ago and start building specialized models

7:55

again. Only now, rather than building classification

7:57

models, we're actually we're still using the

7:59

transformers and the LLLMs, but we're making

8:01

them do specific tasks. specific tasks, use

8:04

this prior knowledge to have a nice

8:06

architecture, a supervised data set for that

8:08

and just do that from scratch. This

8:10

is something that's gonna probably work much

8:12

better, but again you need to make

8:15

sure that the downstream application will never

8:17

go. two out of distribution. So that's

8:19

why it really depends on the application

8:21

and the type of use cases that

8:23

you have. But I think at least

8:26

here it shows that there exists some

8:28

task where an exocan prediction is not

8:30

the answer. And in fact, it's not

8:32

just not the answer, but it's not

8:34

better than random initialization, which is really

8:36

sort of the worst case scenario. Interesting,

8:39

I mean from a fairness and bias

8:41

point of view a lot of people

8:43

say that you know large language models

8:45

are bad in a way because there's

8:47

the dominance of North American cultures and

8:50

so on. But you could also argue

8:52

the converse which is that the good

8:54

thing about them is that they do

8:56

have some awareness of value you know

8:58

so we can fine-tune them to have

9:01

guardrails and to sort of say the

9:03

right thing and so on is that

9:05

harder to do with this approach. Yeah,

9:07

so here because you are in a

9:09

fully supervised setting, you don't have as

9:12

much flexibility to, let's say, change the

9:14

behavior of your model or it will

9:16

have to take the form of supervised

9:18

fine tuning. But because you don't have

9:20

a generative capability, it's certainly restrict the

9:22

type of interaction you have with the

9:25

model and how you can improve it,

9:27

right? Because the output is just, okay,

9:29

is it a good or bad sentiment?

9:31

something that gives you a full answer

9:33

that then you can try to argue

9:36

against and generate a fine tuning that

9:38

I set from is just okay good

9:40

bad and that's it. Another thing is

9:42

training strategy so you know like the

9:44

big players building these LLLMs they have

9:47

lots of internalized knowledge around you know

9:49

even the order in which you train

9:51

the language models everything is important you

9:53

know certainly in the old days of

9:55

like basic models you know you just

9:57

stick a load of data and then

10:00

no one really cares yeah so you

10:02

know now do people need to be

10:04

sort of thinking about the specialized knowledge

10:06

maybe things thinking about curriculum learning and

10:08

all of this kind of stuff? Yeah,

10:11

so this is a good point. So

10:13

we did the paper recently, called the

10:15

Fair Language Model Paradox, where we show

10:17

that when you do this next token

10:19

prediction, because you have some tokens that

10:22

are very low frequency, it's very hard

10:24

to train on them and it takes

10:26

a very long training. So it's very

10:28

wasteful, right? And the problem is that

10:30

because you lose this next token prediction,

10:33

you need to really capture all your

10:35

distribution of tokens, and so you spend

10:37

a lot of time. If the low

10:39

frequency tokens are not useful to solve

10:41

your task, you actually don't need to

10:43

capture it at all. So in turn

10:46

of training dynamics, this is actually a

10:48

much simpler problem in many cases. And

10:50

what we found empirically is that we

10:52

just... out of the box, use typical

10:54

like supervised training, we don't have to

10:57

play with hyperparameter, optimizer, and you have

10:59

very very stable training. So that's one

11:01

thing that could be also interesting for

11:03

future work is to see, is this

11:05

something that is easier to optimize? And

11:08

maybe that's why those like 7 billion

11:10

parameter model can learn and not overfit

11:12

on like 10,000 samples. And then it's

11:14

also bringing other things that maybe this

11:16

on its own. could be a better

11:18

initiation for a next token prediction as

11:21

well. So this is a very open

11:23

up in the air, but maybe you

11:25

could think of a simpler supervised objective

11:27

that would be a better pre-training solution

11:29

that then you can use for next

11:32

token prediction if you wanted to. But

11:34

at least this would be a better

11:36

starting point from random. So you'll almost

11:38

reverse the train. So we've spoken about

11:40

two extremes. So on the one extreme

11:43

we have pre-training and you can like

11:45

use it for any downstream task. And

11:47

on the other extreme we have, you

11:49

know, you start from scratch just with

11:51

one task. Is there an intermediate solution?

11:54

So what if I did this new

11:56

approach but for multitask? Let's say for

11:58

five tasks. Yeah, yeah. So that's a

12:00

great question. So if you really think

12:02

about it in the limit, you could

12:04

formulate a next token, this one or

12:07

not. So in the... extreme case, you

12:09

could just recover an external prediction on

12:11

one end, and on the other end

12:13

you have what we have here, so

12:15

just one task, very coarse, high level,

12:18

predict if it's a good or bad

12:20

sentiment or whatever. So in between you

12:22

have a huge spectrum that you can

12:24

exploit, and if you can find, as

12:26

you said, maybe five very different representative

12:29

tasks, this should be enough to or

12:31

could be enough to learn the representation

12:33

that is as general as possible, and

12:35

then you can use this for maybe

12:37

new... task that comes on the go.

12:39

So I think the research question is

12:42

how to design the minimum amount of

12:44

task so that you have as diverse

12:46

representation as possible. And of course we

12:48

don't want to go to the extreme

12:50

of just doing again next token prediction.

12:53

But this is a very very nice

12:55

research question because If you have this

12:57

spectrum and you can control where you

12:59

want to be, then you can really

13:01

have a per-use case choice. So it's

13:04

not, okay, you are always here or

13:06

always here. Tell me what you want

13:08

to do, how much new tasks you

13:10

expect your model to be exposed to,

13:12

and I tell you where you need

13:15

to be in this spectrum. So this

13:17

could be like very interesting as well.

13:19

Very cool, very cool. It does make

13:21

me think though that. these models understand

13:23

through naive statistical alignment and is it

13:25

possible that the benchmarks we use just

13:28

don't cap you know they the gap

13:30

of understanding that we've lost from moving

13:32

from the pre-trained models isn't being captured.

13:34

Yeah I think because especially in the

13:36

recent years we focus a lot on

13:39

generative methods, all the evaluation and the

13:41

type of objectives we put on ourselves

13:43

is really about good generation, right? Even

13:45

if you want to answer a question,

13:47

you need to generate a good explanation,

13:50

you need to understand what are the

13:52

intermediate steps, and I think the fact

13:54

that we focus on generative models means

13:56

that we completely bias, the evaluation and

13:58

the way we approach this thing, and

14:00

maybe you could have still knowledge that

14:03

is learned without being able to generate

14:05

anything. So I think this is also

14:07

something that could be interesting to look

14:09

at, at least keep in mind when

14:11

we explore those models. But philosophically though,

14:14

isn't generation analogous to thinking in some

14:16

sense? So don't models that generate, aren't

14:18

they smarter in some deep way? Probably

14:20

what you want to do is maybe

14:22

imagine what could be, but I don't

14:25

think you want to do generation is...

14:27

with very granular details like next token

14:29

generation. Because if you think about it,

14:31

even just in terms of like classification

14:33

tasks, you have a lot of different

14:36

uncertainty depending on the token. If I

14:38

start the sentence, okay, I saw this

14:40

movie for minutes, there is no way

14:42

you can tell what was the next

14:44

token for after four, right? So this

14:46

means that you know a period would

14:49

be like a time. component, right? Maybe

14:51

it's like one hour, 10 minutes, two

14:53

hours, but do you really need to

14:55

be able to generate the, I don't

14:57

know, 52 minutes or whatever the answer

15:00

was to actually understand that I was

15:02

seeing a movie, therefore I was staying

15:04

in a place for at least more

15:06

than five seconds, right? So I think

15:08

token is way... to granula. And if

15:11

you had a concept token, that's where

15:13

you could start seeing, okay, this is

15:15

meaningful because that's closer to maybe what

15:17

we do. But right now we are

15:19

very, very, very low level because tokenization

15:21

is a loss less compression, right? So

15:24

this is too close to the raw

15:26

data. And yet, yet we have the

15:28

life easy compared to a computer vision

15:30

because already you work in language, which

15:32

is very compressed representation of knowledge. but

15:35

still token is probably too low level

15:37

still. Well that was a fascinating paper.

15:39

Let's move on to your next one.

15:41

So the birth of self-supervised learning as

15:43

supervised theory and that was with Yam

15:46

Lagoon. Yes. And yeah basically you said

15:48

that the observed differences between self-supervised learning

15:50

and supervised learning are not due to

15:52

the loss function themselves but rather the

15:54

labelling of the data set using training.

15:57

Give us the elevator picture. Yeah so

15:59

basically what we show in this paper

16:01

is that you can have a supervised

16:03

objective like let's say least squares to

16:05

make it simple. the labels and you

16:07

can turn this objective which tries to

16:10

predict sample XN to prediction YN into

16:12

a self-supervisor learning objective which tries to

16:14

compare samples with each other. So basically

16:16

you go from saying okay this image

16:18

is a car or a dog to

16:21

saying are those two images the same

16:23

or not which is like the self-supervised

16:25

type of joint ombudsmaning world. And so

16:27

you can show that If you have

16:29

labels or you have knowledge of this

16:32

perwise relationship, they are actually learning the

16:34

same representation up to some symmetries that

16:36

is irrelevant if you do linear probably.

16:38

So the loss function in itself, the

16:40

SSL one or the supervised one, try

16:42

to do the same thing. They just

16:45

operate on a different view of the

16:47

labeling. whether this image is that or

16:49

are those two images or two samples

16:51

representing the same thing. So given that,

16:53

then the next question is OCOM self-supervised

16:56

learning is able to generalize better than

16:58

supervised. And from this perspective, what you

17:00

can say is that it's because it's

17:02

as if they were solving a supervised

17:04

task, where the labels are not about

17:07

predicting all the cars to cars, but

17:09

are very, very fine grain label, where

17:11

in the limit, each image is its

17:13

own class, basically. If you think about

17:15

supervised learning in this extreme setting, you

17:17

also don't overfit to the task because

17:20

you don't collapse any image to another

17:22

one. And so theoretically speaking, you can

17:24

solve many downstream tasks as you want.

17:26

So this equivalence of loss at least

17:28

brings a slight new perspective on the

17:31

fact that it's not really about the

17:33

objective, it's more about how you design

17:35

the SSL pipeline, or you say, okay,

17:37

with this sample is related to this

17:39

sample, but it's not the objective that...

17:42

makes you learn a better representation. Okay,

17:44

and in the paper you were talking

17:46

about how SSL can maximize the worst

17:48

case downstream task performance. Can you sketch

17:50

that? Yeah, so basically if you think

17:53

about all the possible realization of downstream

17:55

tasks, you could have some very coarse

17:57

scale ones, we have some very coarse

17:59

scale ones, we have maybe different pictures

18:01

of cars and buses, and you just

18:03

want to circuit a car or a

18:06

bus, so no details need to be

18:08

uncoded to solve this. But then you

18:10

can have downstream tasks where you want

18:12

to solve or you. So the point

18:14

now is that you want to learn

18:17

a representation, so that if you look

18:19

at the distribution of downstream tasks performance,

18:21

you are able to be as good

18:23

as possible on most of them. So

18:25

you don't want to be very good

18:28

on some and then in the tail

18:30

you are very bad on the majority

18:32

of them. And so then from this

18:34

you can try to say okay what

18:36

will be the labelling that tries to

18:38

make your worst case as good as

18:41

possible and from this you can say

18:43

okay this is actually the labelling that

18:45

self-supervised is actually implicitly doing. How does

18:47

the class balance affect the difference in

18:49

the losses? Oh yeah so this is

18:52

a very good point actually in a

18:54

follow-up paper we are doing right. right

18:56

now, we show that current SSL objective

18:58

assume class balanceness. And this is something

19:00

we already highlighted quickly in this. as

19:03

self-supporting as a uniform cluster prior to

19:05

paper, we did a couple years ago,

19:07

and we show that current assessor objectives

19:09

assume balance representation of classes or concepts.

19:11

And this means that if you train

19:14

on ImageNet, things work out very well

19:16

because concepts are sort of equally represented.

19:18

But then if you go to other

19:20

data sets like I Naturalist, which are

19:22

very avi-tail, then you have a huge

19:24

bias in your representation. So until now,

19:27

people do not really know how to

19:29

solve this. One way people approach this

19:31

is through data curation. And they say,

19:33

OK, I'm just going to remove the

19:35

oversampled concepts to try to make it

19:38

more uniform. And then I do self-supervised

19:40

learning on this. But because now we

19:42

have this theoretical formulation and this equivalence

19:44

of losses, we can use the exact

19:46

same settings that people used in a

19:49

supervised learning to re-weight depending on the

19:51

frequency of classes. We can use that

19:53

to come up with a new self-supervised

19:55

learning loss that takes this imbalance into

19:57

account. This type of thing is enabled

19:59

from this mathematical formulation and its principle.

20:02

So the way we do this waiting,

20:04

you can prove that it is the

20:06

right way to do it from this

20:08

supervisory. And so this is really nice

20:10

because suddenly from this seemingly naive connection,

20:13

you cannot come up with new generation

20:15

of self-supervised learning models where you can...

20:17

actually match what the real world data

20:19

distribution is like. So non-uniform distribution of

20:21

classes, maybe even if you have some

20:24

samples that are more noises and others,

20:26

you can include that information as part

20:28

of the SSO objective as well. So

20:30

suddenly you have a new world of

20:32

possibilities that comes and because there is

20:35

this connection, you can actually prove, okay,

20:37

this is the right way to do

20:39

it, at least from this supervised theory

20:41

viewpoint. You also pointed out a connection

20:43

to V Craig. Exactly. So basically, what

20:45

we do in the paper that we

20:48

show if... you have a least square

20:50

supervised type of objective and you turn

20:52

it into a SSL one, what you

20:54

obtain is basically V Craig. So then

20:56

you have a few variations, it could

20:59

be V Craig or WMC, depending on

21:01

how you do this from supervised to

21:03

SSL, but you can show that depending

21:05

on the type of supervised loss, you

21:07

recover different type of SSL ones. If

21:10

you look maybe more at cross-anthropy, a

21:12

super-wise learning is going to be more

21:14

like a simpler type of loss, but

21:16

you have this one-to-to-to-one correspondence correspondence. And

21:18

this is also very nice. And this

21:20

is also very nice. because in supervised

21:23

learning at least you know when one

21:25

loss may be preferred compared to another

21:27

one and this has been studied for

21:29

a long time right because supervised learning

21:31

has been around forever and so now

21:34

we can we use those insights for

21:36

self-supervised learning. So this to me is

21:38

also a very very strong benefit of

21:40

this thing is that suddenly all the

21:42

theory and like the thousands of papers

21:45

that have been done in supervised learning

21:47

we can just take it and apply

21:49

it in SSL. Another example is a

21:51

neural collapse for example that has been

21:53

proven in supervised setting. Now it applies

21:56

like in five lines in a SSL

21:58

setting as well. So this connection is

22:00

really... beyond just trying to say, okay,

22:02

it's not the objectives that make SSL

22:04

better. It's really tying those two huge

22:06

communities together towards the goal where you

22:09

have a single unified objective to learn

22:11

representation. And this is nice too, because

22:13

if you speak to people, they will

22:15

think, okay, you have super-wise learning on

22:17

one side, SSL on the other side,

22:20

and basically you are either in one

22:22

camp or the other. But now what

22:24

we show is that you actually SSL

22:26

is... pretty much everything in representation learning

22:28

and supervise just one realization of SSL.

22:31

Then V Craig without labels, he knows

22:33

other ones, then this one is another

22:35

one. So you really have a better

22:37

understanding of this relationship and what representation

22:39

learning is trying to do. Galaxy brain

22:41

question incoming. Could you combine SSL and

22:44

supervised objectives in some way to improve

22:46

generalization? Yes, yes. So there is one

22:48

paper which is a supervised contrastive learning.

22:50

So the way they do it is

22:52

that they use the labels within a

22:55

simpler framework to try to basically do

22:57

fully supervise learning, but with a simpler

22:59

objective. So first of all, we can

23:01

show that indeed this makes sense and

23:03

that basically we can explain the empirical

23:06

result that they have. But actually we

23:08

can do a little bit more than

23:10

So if you are in a semi-supervised

23:12

setting, for example, it may not be

23:14

clear how to combine those two losses

23:17

anymore, or maybe you could say, you

23:19

have the two and have a coefficient

23:21

to weight them, but then you need

23:23

to cross-validation and so on. But now

23:25

from this perspective, you can combine them

23:27

in a very principled way, and you

23:30

can understand which weighting makes sense depending

23:32

on... how much sample you have in

23:34

one or the other. And you can

23:36

use all the literature, again, from supervised

23:38

learning, for this setting as well. So

23:41

this is something we can do very

23:43

easily with this formulation as well. Okay,

23:45

so if SSL and Supervis are two

23:47

sides of the same coin, I mean,

23:49

of course, we can use this theoretical

23:52

framework to design new forms of SSL

23:54

framework. Does it, but you know, is

23:56

the distinction relevant if they are the

23:58

same thing? I think it's not just

24:00

two sides of the same coin. SSL

24:02

is more general than supervised learning. So

24:05

it's really, SSL could be the more

24:07

general objective to learn representation. The more

24:09

prior knowledge you have, the more you

24:11

know about your labels and then SSL.

24:13

like slowly becomes supervised learning through the

24:16

labels that you use for the SSL

24:18

objective. But then, because as you said,

24:20

you have this hierarchy, now, it does

24:22

not really make sense to say you

24:24

have either supervised learning or SSL. Rather,

24:27

what makes sense is you say, okay.

24:29

what this relation matrix, what is per

24:31

wise matrix? If you build it from

24:33

labels, it's a supervised learning, if you

24:35

build it from other a prior knowledge,

24:38

for example, two consecutive frames in a

24:40

video, I basically have the same class,

24:42

then you are more in a unsupervised

24:44

SSL setting, but it's all about how

24:46

do you build this per wise relation

24:48

matrix? that's the main question. Very cool,

24:51

right, let's move on to your next

24:53

paper. So no location left behind, measuring

24:55

and improving the fairness of implicit representations

24:57

for Earth data. So there's loads and

24:59

loads of modeling frameworks now that do

25:02

these implicit neural representations of geospatial Earth

25:04

data. So things like, climate modeling, resource

25:06

allocation, environmental modeling. I was actually interviewing

25:08

Johannes from NxAI, yes, I don't know

25:10

if you know him, but he's working

25:13

on similar stuff. Okay. So basically what

25:15

we show is that when you want

25:17

to model, for example, let's say temperature

25:19

or precipitation to make it simple, and

25:21

you want to learn, for example, implicit

25:23

neural representation, it means that you want

25:26

a model so that if you give

25:28

a location and a... date, for example,

25:30

it can predict what was the temperature

25:32

there. So if you have this type

25:34

of implicit neural representation, it's very good,

25:37

because if you learn a nice model,

25:39

then you can actually interpolate those values,

25:41

so maybe estimate what the temperature was

25:43

in this part of the globe, or

25:45

you do not have a sensor. But

25:48

you can also do extra polation as

25:50

well. If you assume you really learn

25:52

the true physical model of the world,

25:54

you could say, okay, what the temperature

25:56

will be two years from now, right?

25:58

for all sorts of applications. The thing

26:01

is that when you do this nowadays,

26:03

depending on the architecture and the different

26:05

design choices that you do, you will

26:07

maybe have a very good prediction on

26:09

average, so when you look at the

26:12

rate performance around the world globe, but

26:14

actually if you look for example around

26:16

islands or coastal area, your prediction is

26:18

going to be very bad almost random.

26:20

So this is something that can be

26:23

very concerning because if you use this

26:25

type of model to decide about a

26:27

policy that will affect a specific island.

26:29

Using this model prediction is as good

26:31

as using random guesses. So it can

26:34

be very detrimental and people need to

26:36

be aware of those biases. So what

26:38

we found is that, for example, for

26:40

this type of climate data, islands are

26:42

often disregarded, coastal area, basically region where

26:44

you have a big gradient in the

26:47

type of data that you try to

26:49

model. How much of a responsibility do

26:51

modelers have to detect these kinds of

26:53

biases in the data? So I think

26:55

there is like two components, as you

26:58

said. So one could be that just

27:00

the dynamic of the data you are

27:02

trying to model is harder near island

27:04

or maybe it's even unpredictable because you

27:06

don't have enough observations to do that.

27:09

So you have some uncertainties that probably

27:11

you can never recover from good design.

27:13

But still what we found here is

27:15

that a lot of the biases now.

27:17

comes from the architecture and all you

27:19

want to do to encode those positions,

27:22

the type of basis you use to

27:24

do the prediction. So right now it

27:26

seems that a big chunk of the

27:28

bias... from the architecture, but I totally

27:30

agree that I don't think we can

27:33

remove the bias entirely because there is

27:35

maybe just different type of uncertainty, a

27:37

different part of the planet as well.

27:39

I mean the world is a very

27:41

very complicated place. I mean realistically to

27:44

what extent can we mathematically model it?

27:46

Yeah so that's a good question. So

27:48

I think it depends the type of

27:50

horizon that you have and the type

27:52

of data that you want to model.

27:55

If you have a system that is

27:57

much more chaotic or can vary very

27:59

quickly without... much changes in the past

28:01

observation. That's something that current models are

28:03

having a very hard time with. If

28:05

you want to predict something else, for

28:08

example... temperature in North America, not near

28:10

the coastal area, so really inland, maybe

28:12

that's where you have less, gradient dynamics,

28:14

things are a bit more stationary, especially

28:16

in through time, so then it can

28:19

become much better. But I think at

28:21

this point, we don't have an architecture

28:23

that is really able to understand that

28:25

you have different physics, different dynamics models

28:27

at different parts of the globe. And

28:30

so because of this, you just... see

28:32

what's the best on average and it

28:34

means you miss out a lot of

28:36

details. Can you tell us about some

28:38

of the technical framework? So one thing

28:40

we showed, for example, at least for

28:43

this type of globe data representation, is

28:45

that people use a four-year basis to

28:47

model the prediction. And this is something

28:49

that is better than not using any

28:51

basis at all. But what it means

28:54

that you imply the type of signal

28:56

you're predicting is very stationary and not

28:58

localized at all. And this is a

29:00

very strong prior, right? So this may

29:02

be true for some things, but for

29:05

other things like precipitation precipitation or... temperature

29:07

where you have localized very high gradients

29:09

then it's a strong bias and if

29:11

you come from signal processing community you

29:13

know very well that to have better

29:16

localization you go from four years to

29:18

wavelets and so that's one thing we

29:20

did in this paper and we show

29:22

that using wavelet bases to encod those

29:24

data allows you to have better localization

29:26

and this removes of the biases. And

29:29

here it's more of proof of concept

29:31

that different design choices give you a

29:33

different type of bias trade-off. Wevelette is

29:35

not the answer to everything, right? But

29:37

I think the next step is to

29:40

really be able to encode less and

29:42

less a priori which basis to use

29:44

and let the model learn from the

29:46

data on its own. And we are

29:48

not yet at this point, at least

29:51

for this type of climate data. How

29:53

could it handle noisy or missing data?

29:55

This depends really on the type of

29:57

model you use. So for example, if

29:59

you have INR, then you will not

30:01

choose the missing data as part of

30:04

your training pipeline and that's one of

30:06

the benefit of them. So if one

30:08

of your sensors stopped recording during some

30:10

years, you just don't choose that as

30:12

part of your training data because you

30:15

really control where do you have the

30:17

data and when you have it, what

30:19

the prediction should be. So these earth

30:21

models, they are now informing policy around

30:23

the world. Who should we hold accountable?

30:26

I mean, is it the technology, is

30:28

it the scientists who design the models,

30:30

is it the policy makers who interpret

30:32

the results? I think it's very hard

30:34

for the person who designs the model

30:37

to know priori what is going to

30:39

be used for. So I think it's

30:41

more downstream. When you know clearly what

30:43

you want to do with it, you

30:45

should first set up a nice evaluation

30:47

pipeline to make sure that it's something

30:50

you can actually use to make those

30:52

decisions. And then you can report any

30:54

type of... your model for people to

30:56

improve on the design, but prior to

30:58

it's very hard to imagine what this

31:01

model will be used for. So in

31:03

an ideal setting you wish that there

31:05

would be no bias at all, but

31:07

in practice the world of possibilities being

31:09

so large it needs to be more

31:12

of a feedback loop and then... iterate

31:14

until you have something that you can

31:16

really trust and then you can act

31:18

on it. Modeling data is very anthropocentric,

31:20

right? So, you know, we focus on

31:22

human populations and so on. Should we

31:25

also focus on, you know, like just

31:27

ecosystems and places that have got nothing

31:29

to do with humans? Oh yeah, that's

31:31

a great question and in fact that's

31:33

one of the big issue with... a

31:36

lot of the data set which is

31:38

a crowd source set because by definition

31:40

the amount of data that you get

31:42

is proportional to the number of users

31:44

you have depending on the location. And

31:47

this means you have a huge bias

31:49

in what your model is learning and

31:51

what your model is focusing on, which

31:53

means you miss out on a lot

31:55

of things. So I think that's also

31:58

one thing that, okay, crowdsourcing can give

32:00

you a lot of data quickly, but

32:02

it's very biased data. So then the

32:04

question is, how much of this bias

32:06

data versus maybe paying a lot more

32:08

and capturing other part of the two

32:11

you should have? And maybe you could

32:13

be able to show that under some

32:15

specific condition, just having 10% of the

32:17

data, which is high quality, uniformly sample,

32:19

and then 90% which is crowd sources,

32:22

you can try to use those 10%

32:24

to incur your representation and then use

32:26

all that data together. But there is

32:28

a huge amount of research question in

32:30

that, because that's a very big source

32:33

of bias. And there's a bit of

32:35

a policy question, but we are using

32:37

these things, you know, to do resource

32:39

allocation, right? giving more resources to some

32:41

populations might be taking it away from

32:43

others and then there's the fairness over

32:46

time thing as well which is that

32:48

what is fair like now might not

32:50

be fair in a hundred years time

32:52

so how should we think about it?

32:54

Yeah that's a good question I think

32:57

this is also very application. If you

32:59

want to predict where to build a

33:01

house to solve some specific problem, maybe

33:03

you don't really mind having bad prediction

33:05

where there is no population anyway because

33:08

you are not going to build a

33:10

house there. So in this case, maybe

33:12

the crowdsourcing type of data is actually

33:14

good, but this could really be dependent

33:16

on the type of application. And just

33:19

one thing I would say regarding the

33:21

point you made before, this type of

33:23

bias actually is something that you have

33:25

in computer vision. So there is like

33:27

a very nice... paper done by a

33:29

Marquis brain. Basically, they showed that most

33:32

of the data we have, like from

33:34

ImageNet, is from North America. And so

33:36

maybe you reach like 90% state of

33:38

the art performance to predict, for example,

33:40

type of chairs. but only

33:43

for North American models.

33:45

And when you

33:47

start looking at type

33:49

of cars or

33:51

chairs cars or Africa or

33:54

East Asia, the model

33:56

model performance is extremely

33:58

bad. So type of

34:00

problem is something you

34:02

have across across modalities

34:04

that's something that's very big,

34:07

big issue. it's it's

34:09

always a pleasure

34:11

and an honor to

34:13

have you on

34:15

the show. the you

34:18

so much. you so much.

34:20

Thank you so much.

34:22

thank you so much.